[Music] today our focus is on artificial intelligence well actually i should say machine learning a certain type of ai in this video i want to introduce the concept of the machine learning value chain this will help us to think about machine learning as a product and we can also use this framework to get an idea of the value of data a key question we will ask is can more data help to make better decisions related to that we'll think about whether data and algorithms are complements or substitutes and we will see that the answers matter
quite a bit for using ai in a business context the machine learning value chain consists of four components algorithms or models data computation or training and insight or innovation the value chain can be linear in a sense that you apply algorithms to data using computers to arrive at predictions we can also have more complex interactions with loops between components but let's not go there yet let's first think about who is producing controlling and owning the components of the machine learning value chain let's start with algorithms and models who is investing in the research and development
that goes into them who is controlling the intellectual property it turns out that this industry is largely occupied by non-commercial organizations such as research universities that produce basic research and fundamental technologies most knowledge about advances in machine learning is published in academic papers but there's also an industry of commercial organizations that offer ready-to-use algorithms such as ibm google microsoft and a large number of startups finally downstream firms that apply algorithms to their business problems are also investing in in-house capabilities in specialized human capital just like in other forms of research and development what about data
who is investing in the collection and quality of data who owns data we often see that downstream firms own proprietary data sets such as customer databases or data from sensors in devices and machines with the machine learning value chain data can create a competitive advantage if your data lets you observe something that your competitors do not observe for instance say that amazon has better information about consumer preferences than google because google can only observe what you search for whereas amazon knows what you end up buying but of course there are also upstream markets for data
a sizeable industry of data brokers is in the business of selling individual level information about consumption preferences or financial health market research firms specialize in tracking industry-wide numbers related to advertising spending demand indicators and so on the competitive advantage of such firms is often that they get to observe an entire industry not just the information that any individual firm would have about their own ad spending and sales okay let's now talk about computation or training how does this industry look like a few firms specialize in the research and development of hardware and new categories of
hardware have been developed and optimized for machine learning from the perspective of downstream firms you can either go and buy these hardware components and build your own firm specific infrastructure or you can rent data center infrastructure by using third-party cloud computing services this industry appears to be concentrated with a few dominating firms and what the suppliers often do is bundle cloud computing services with ready-to-use algorithms with algorithms data and computing power we hopefully arrive at insights this stage of the machine learning value chain is all about firm specific decision making based on predictions from decision
support systems insights from machine learning can be used for process innovation for example by making things cheaper or more productive and for product innovation for example by enabling new products or product features either way there's the potential for machine learning to yield insights and innovation for both upstream firms that produce components of the value chain or downstream firms that apply the components of the value chain speaking of applying the components of the machine learning value chain how can we think of machine learning as a product if we put the whole value chain together well the
first observation is that algorithms and software implementations are mostly free but the computing power meaning access to hardware is not free data is also not free it needs to be collected combined and sometimes labeled but as mentioned already it can represent the competitive advantage one last observation is that along the value chain firms operate on a range of business models there are firms selling complementary goods such as computing power some cell combined packages such as pre-trained algorithms and consulting services others try platform business models where they either provide algorithms data and or computing infrastructure to
help incentivize innovation from complementors now that we have established the four components of the machine learning value chain and have thought about business models and markets around them let's go into a bit more detail we will discuss fundamental types of algorithms and how they interact with data this will help us think through circumstances where the machine learning value chain is not linear but has more complicated loops there are two broad types of machine learning supervised and unsupervised algorithms what is supervised machine learning quite simple you have training data with known labels for example you know
whether a picture shows a cat or a dog if you're familiar with econometrics think of a regression model where you know the outcome and a number of explanatory variables we know whether it's a cat or a doc and we can correlate that with information on the colors of thousands of pixels contrast that to unsupervised machine learning where you do not observe labels you only observe the features on the right hand side of your regression equation and you want to find similarities among them for example you have a bunch of pictures of animals but you do
not know whether they are cats or dogs or giraffes unsupervised machine learning will help you to learn that everything that has brown hair a short neck and two ears is more likely to be a cat or a dog than a giraffe the key difference between supervised and unsupervised machine learning is the data that they need enabling data is costly what is common for both types of models is their purpose we want to predict stuff machine learning is very good at that but you can't learn much about causality in stable settings this is totally fine for
example it's fine to not understand why cats look like they do as long as we have figured out the correlations and predictions based on those work well for either type of machine learning algorithm data is important but how much data do we need to arrive at acceptable predictions or zooming out to see the big picture again under which circumstances can more data help to make better decisions actually what we should be asking is what is the optimal level of data optimal can be defined in a variety of ways but let's for now focus on firm
performance and profits for instance since data is not free simply adding more and more data will not be profit maximizing at some point the additional insights you generate from adding more data are not as valuable compared to how much it costs to collect or acquire more data data can be characterized according to quantity and quality the first aspect of quantity refers to features the number of variables that you can observe for example you know my name where i live where i work my phone number my email address and so on but then there's also the
aspect of observations the number of cross sections for instance you do not only have the data about me but also about my wife the third dimension of quantity relates to repeated observations over time for example you know the location of my cell phone not only once but in intervals of 15 minutes finally how can you think about data quality apart from obvious issues of factually wrong information or mismeasured data from sensors an important dimension of data quality is variation more data does not necessarily mean that it is more useful if two people have the same
characteristics having data about one of them is enough in many applications without variation adding more data is not very helpful so back to the question can more data help to make better decisions what happens if we increase data quantity we reduce what econometricians call omitted variable bias and it also helps to predict rare events simply because rare events are more likely to be included in a larger sample in addition with repeated observations you can subtract unobserved time-invariant differences between individuals in econometrics terms more data can help you to correct for bias from unobserved heterogeneity can
better data help to make better predictions certainly yes higher quality data helps a lot with measurement error issues endogenetic issues and so on a very good source of variation are experiments for example handing out coupons to a random selection of customers will generate variation in purchase decisions and will help you figure out optimal pricing strategies increasing data quantity and quality is not free making better decisions with more and better data is always a trade-off between costs and benefits collecting repeated observations takes time running experiments has direct costs opportunity costs and potential externalities what makes this
trade-off especially interesting and difficult to navigate is that you need to compare costs and benefits ex-ante although it's kind of the point of using prediction technology that you don't know the benefits yet finally let's talk about another dimension the relationship between algorithms and data are models and data complements or substitutes the answer seems obvious models and data are complements in the sense that they have zero independent value for instance what is an algorithm if you cannot apply it to data but then isn't an important reason why we use machine learning that we want to predict
data points that they do not have doesn't that mean that models and data can be substitutes well yes the circumstances under which more or better data can substitute for better models or the circumstances where better models can substitute for more or better data will depend a lot on the firm specific context why is it important to know whether and when models and data are complements or substitutes well it matters a lot for firm strategy and in which components of the machine learning value chain to invest in right it also matters a lot for the choice
of organization for the interactions downstream firms will have with firms in upstream markets and finally looking at the big picture whether models and data are complements or substitutes matters a lot for policy for example think of the trade-off for governments to regulate consumer privacy and or provide funding for research and development of ai through public universities to sum up in this video we've talked about the four components of the machine learning value chain we use this to think about machine learning as a product we've also looked at the relationship between data quantity and quality and
business decisions finally we determined why it matters whether models and data are complements or substitutes
